Boundary thickness and robustness in learning models